Pith. sign in

REVIEW 4 major objections 6 minor 126 references

CReSL claims graph domain adaptation fails when class evidence shifts across neighborhood ranges, and fixes it by learning a soft source-to-target resolution routing from class prototypes.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 08:24 UTC pith:VZCVS6YU

load-bearing objection A genuinely new mechanism for GDA that deserves referee time, but the empirical case is under-supported by missing tables, no code/data, and a calibration loop that could collapse. the 4 major comments →

arxiv 2607.29365 v1 pith:VZCVS6YU submitted 2026-07-31 cs.LG

Cross-Resolution Semantic Learning for Graph Domain Adaptation

Yingxu Wang , Haoze Huang , Zhongkai Zheng , Shangsong Liang This is my paper
classification cs.LG MSC 68T0768R10
keywords graph domain adaptationsemantic resolution shiftcross-resolution routingprototype transporttarget graftingmulti-resolution graph representationgraph classificationdistribution shift
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Graph domain adaptation has implicitly assumed that the neighborhood range carrying class evidence is the same in source and target; CReSL argues this is often false, and that this mismatch—called semantic resolution shift—is a hidden cause of negative transfer. The paper proposes learning a soft routing from each source resolution to target resolutions, derived purely from cross-domain class structure: target posteriors and source labels form class-resolution prototypes, prototype discrepancies are converted into a refined routing matrix, and that routing guides both prototype matching and instance-level target grafting. The reported experiments on twelve graph-classification shift settings show CReSL outperforming representative baselines in most cases, and a post-hoc oracle analysis finds the learned routing picks the same target resolutions as an accuracy-based oracle for all four source experts. A target-risk decomposition (Theorem 1) backs the design, bounding target risk by source risk plus the two routing-weighted terms. The paper is trying to establish that resolution correspondence is not a nuisance but a learnable structure that should be part of the GDA objective.

Core claim

The central claim: under distribution shift the propagation resolution at which class-discriminative evidence is strongest can change between domains, so the default pairing of each source resolution with the same target resolution is a design flaw, not a harmless simplification. CReSL replaces that default with a label-free, learnable correspondence. A shared GNN with learnable resolution embeddings builds comparable representations at J+1 propagation orders, each with a resolution-indexed expert. A row-stochastic routing matrix Γ is then iterated: current Γ yields soft target posteriors; posteriors and source labels define class-resolution prototypes in both domains; the class-averaged pro

What carries the argument

The load-bearing object is the routing matrix Γ ∈ R^(J+1)×(J+1): a row-stochastic table whose entry Γ_{j,k} controls how much source-resolution expert h_j is trusted on target resolution k. It is updated in a closed loop. Current Γ aggregates expert responses over target resolutions into soft target posteriors (Eq. 9); posteriors and source labels yield class-resolution prototypes (Eq. 10); the class-averaged prototype discrepancy D (Eq. 11) is converted by row-wise softmax into the refined routing Γ+ (Eq. 12); and Γ+ weights both the prototype-transport loss (Eq. 13) and the target-grafting consistency loss (Eq. 17), then becomes Γ for the next step. This loop is what converts raw class str

Load-bearing premise

The method's gains rest on the soft target posteriors being class-calibrated from the start: if the early posteriors are wrong, the prototype discrepancies they induce push the routing and grafting in the wrong direction, amplifying rather than correcting the initial error.

What would settle it

Three checks would settle the claim. First, on a second benchmark, compare the learned routing Γ against a post-hoc oracle matrix (source-expert accuracy ranked over target resolutions): if the top-1 agreement collapses well below the reported 4/4, the label-free correspondence is dataset-specific. Second, initialize Γ with a deliberately wrong order (e.g., reversed resolutions) and run the full loop: if the routing stays wrong or accuracy collapses, the coupled estimation is fragile rather than self-correcting. Third, engineer a shift with no resolution mismatch (same informative range in bot

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Same-resolution pairing is a hidden defect: ablations show replacing learned cross-resolution routing with fixed same-resolution pairing consistently degrades accuracy, so resolution correspondence should be an explicit adaptation variable, not a default.
  • The correspondence can be learned without target labels: CRPT builds class-resolution prototypes from source labels and soft target posteriors, and the refined routing matches the oracle-preferred target resolution 4/4 on the Mutagenicity analysis.
  • Posterior-weighted grafting beats hard assignment: mixing class-conditioned displacements by class probabilities outperforms committing each target graph to its maximum-posterior class.
  • The target-risk bound ties theory to the objectives: minimizing source risk, prototype discrepancy, and grafting inconsistency jointly controls target risk up to residual terms.
  • The gains appear across both structural and feature domain shifts on chemical and protein benchmarks, not just one shift type.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A testable failure mode follows from the coupled loop Γ → posteriors → prototypes → Γ+: if early target posteriors are miscalibrated, prototype discrepancies may lock in the wrong routing; initializing Γ with a deliberately reversed permutation would reveal whether the loop self-corrects or amplifies.
  • The machinery is modality-agnostic: any representation family indexed by an ordered scale—context window, kernel width, layer depth—could be routed by the same prototype-transport loop, though the paper only demonstrates graphs.
  • The oracle-comparison diagnostic generalizes: building the same post-hoc accuracy oracle for any GDA method would quantify how much resolution mismatch actually exists in a given shift, pushing benchmark design toward per-resolution diagnostics instead of aggregate accuracy.
  • If the claim holds broadly, fixed same-resolution alignment in existing GDA methods is not merely suboptimal but a systematic negative-transfer risk whenever the shift changes the informative range.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper introduces CReSL, a graph domain adaptation (GDA) method that explicitly models cross-domain shifts in 'propagation resolution'—the neighborhood range over which class-discriminative evidence is expressed. It builds a multi-resolution representation bank using a shared GNN plus learnable resolution embeddings, attaches a linear expert to each source resolution, and then (i) computes soft target posteriors from a routing matrix Γ (Eq. 9), (ii) updates Γ via class-resolution prototype discrepancies (Eqs. 10–12), and (iii) adapts target representations by posterior-weighted prototype displacements and routing-weighted KL consistency (Eqs. 14–17). The total objective is Eq. (20). A target-risk decomposition is stated as Theorem 1. Experiments on Mutagenicity and six feature-shift datasets report accuracy in Table 1; ablations and sensitivity analyses on Mutagenicity appear in Table 2 and Figures 4–7.

Significance. If the empirical results are reproducible, the idea of explicitly learning soft source-to-target resolution correspondence from class-conditional geometry is a meaningful addition to GDA. The oracle-compatibility analysis (Figure 5) is a useful check: CReSL's learned Γ agrees with oracle top-1 choices for all four source experts on Mutagenicity. The ablations support the usefulness of each component, and the paper states a theoretical decomposition. However, the current manuscript is incomplete as an archival submission: the main text refers to absent tables and to ogbg-molhiv results, no code or data are released, and no significance tests are reported. The central empirical claim therefore rests on one quantitative table and cannot yet be fully verified. The self-training/coupled-estimation mechanism also lacks a stability analysis, which is a substantive correctness risk.

major comments (4)
  1. [Experiments, 'Performance Comparison'] The text states that results are reported in 'Tables 1 and ??–??' and that performance is evaluated under 'diverse domain shifts', but the only quantitative table is Table 1. Moreover, ogbg-molhiv is listed among the datasets, yet no results for it appear anywhere. The abstract's claim of outperforming baselines 'across most settings' is load-bearing; the current manuscript supports it with only 12 transfer settings and no table for the claimed additional experiments. This must be fixed by supplying the missing tables/results or by scaling back the claims.
  2. [Eqs. (9)–(12), CRPT] The central mechanism is a coupled estimation loop: target posteriors in Eq. (9) are computed from the current routing Γ, while the updated routing Γ+ in Eq. (12) is built from those same posteriors via prototypes. If early posteriors are miscalibrated, the prototype discrepancies in Eq. (11) reinforce incorrect routing and grafting directions. The paper acknowledges this coupling in Figure 2 but introduces no confidence gating, entropy regularization, or source-side cross-resolution consistency loss. In addition, each expert h_j is trained only on source representations at its own resolution j (Eqs. 6 and 18), so evaluating h_j on target resolution k≠j is out-of-distribution for the expert. Theorem 1 does not remove the concern because its residual term ΛΘ is left uncontrolled. Please add an explicit stability analysis or a mechanism to prevent error amplification, and report sensitivit
  3. [Theoretical Analysis, Eq. (21)] The theorem's bound contains the term ΛΘ, which is described only as 'residual prior, conditional-shape, posterior/prototype, and expert-mixture errors' and is never explicitly defined. Without a concrete expression or bound on ΛΘ, the statement 'minimizing these terms tightens the bound' is not formally justified. The caveat 'when the residual terms are sufficiently small' is not a checkable condition. Please state ΛΘ explicitly and give conditions under which it is bounded or negligible.
  4. [Ablation Study, Table 2] The ablation results are reported without standard deviations, and no significance tests are given. Some differences between variants are nontrivial (e.g., M0→M1: CReSL 80.7 vs. 75.9 for w/o CRTG), but several are modest (e.g., M3→M0: 63.4 vs. 60.7 for w/ HG). Since the paper's claim that each component is essential rests on this table, the absence of variance information and significance testing weakens the evidence. Please report run-level variation and, where feasible, paired statistical tests.
minor comments (6)
  1. [Methodology, Eq. (3)] The notation S_G^{r_j} should specify the convention for r_0=0 (presumably S_G^0 = I). This is implicit from the text but worth making explicit to avoid confusion.
  2. [Methodology, Eqs. (6) and (18)] The linear expert h_j is defined on R^d, but in Eq. (18) the source logits are the α-weighted mixture over all j. Clarify that the same expert weights are used for the target posteriors in Eq. (9); this is stated in the text but the notation could be more explicit.
  3. [Experiments, Table 1] The WL subtree baseline is reported without standard deviations. If this is because the kernel method is deterministic, say so; otherwise provide the same run statistics as for the other baselines.
  4. [Experiments, Figure 4] The caption reads 'Class 1 Source PointsClass 0 Target Points'; the missing separator makes the figure difficult to parse. Please rephrase.
  5. [Implementation Details] No random seeds are reported, and no information is given about how the five independent runs vary across seeds. Adding seed ranges would improve reproducibility.
  6. [Overall manuscript] The LaTeX artifacts 'Tables ??–??' and the missing ogbg-molhiv results should be fixed. The paper would also benefit from a short algorithm pseudocode for the alternating Γ/posterior updates, since the coupled estimation is central to the method.

Circularity Check

0 steps flagged

No significant circularity; the learned routing is a self-training fixed point with an acknowledged coupled-estimation loop, but the headline empirical claim is evaluated against held-out target labels and external baselines.

full rationale

The derivation chain does not reduce any claimed prediction to its inputs by construction. The routing matrix Γ in Eq. (9) is used to form soft target posteriors, which in turn define prototypes in Eq. (10) and the refined routing Γ+ in Eq. (12); this is a self-training/coupled-estimation loop that the paper explicitly acknowledges (Introduction, Challenge (2), Figure 2: 'the target posteriors required to estimate this organization also depend on the unknown correspondence'). A self-training fixed point is a stability concern, not a logical circularity, because the evaluated quantity is true target accuracy, not the fitted routing itself. The oracle comparison in Figure 5 is a post-hoc external check ('Target labels are not used for hyperparameter tuning and only for post-hoc evaluation'), and the 4/4 agreement is an empirical result rather than an identity: oracle O_{j,k} is expert accuracy on resolution-k target representations, while Γ_{j,k} is derived from class-averaged prototype distances, which are different functionals. The ablations and comparisons against external baselines (Table 1) are self-contained. No load-bearing self-citation or imported uniqueness theorem is used; the 'propagation resolution' notion is taken from a non-author citation (Pilavcı et al. 2024), and no ansatz is smuggled in via the authors' own prior work. Theorem 1's uncontrolled residual ΛΘ is a limitation but not a circularity. I also note, as a verification/completeness issue rather than circularity, that the text refers to 'Tables ??–??' and to ogbg-molhiv results that do not appear in the manuscript.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 0 invented entities

No new physical entities are postulated. 'Semantic resolution shift' is a conceptual re-framing of representation shift, and the 'experts', routing matrix, and prototypes are model components rather than invented entities. The central empirical claim rests mainly on a small number of hand-chosen hyperparameters and the ad hoc assumption that prototype geometry in a shared latent space reveals cross-resolution compatibility.

free parameters (6)
  • Resolution depth set and pre-propagation orders R = J=3, R=(0,1,2,4)
    Selected based on sensitivity analysis on Mutagenicity (Figure 7a), where performance peaks at J=3.
  • Grafting strength β = 0.5
    Chosen from the sensitivity curve on target-domain accuracy (Figure 7b), yet described as 'pre-specified and fixed across all tasks'.
  • CRPT and CRTG loss weights = λ_CRPT=1e-2, λ_CRTG=1e-3
    Set by hand in Implementation Details; no sensitivity analysis is reported for these weights.
  • Routing temperature τ = not reported
    Appears in Eq. (12) as a positive scalar controlling routing concentration, but no value is given anywhere in the manuscript, making exact reproduction impossible.
  • Resolution embeddings E_res and expert mixture scores a_j = learned during training
    These are learned parameters central to preserving resolution identity and weighting source experts; they are fitted to the source and target data.
  • Architectural dimensions = d=128, d_r=16, dropout 0.2, lr 5e-4
    Chosen by the authors and required to reproduce the reported results; no search is described.
axioms (6)
  • domain assumption Source and target share the same label space Y and source labels are reliable.
    Stated in the Methodology setup and used in Eq. (10) for source prototypes. Standard for GDA but load-bearing.
  • ad hoc to paper Soft target posteriors p_t from Eq. (9) are informative enough to estimate target class structure at each resolution.
    Used to compute target prototypes in Eq. (10). If initial posteriors are wrong, the routing and grafting loop can reinforce errors.
  • ad hoc to paper Euclidean distance between class-resolution prototypes in the shared latent space is a valid proxy for source-to-target resolution compatibility.
    Used in Eq. (11) to define D_{j,k}. No external evidence is provided that prototype geometry encodes compatibility.
  • domain assumption A single shared GNN with added resolution embeddings produces representations that are comparable across resolutions and domains.
    Assumed in Eq. (4) and needed for cross-resolution expert evaluation and prototype displacements.
  • standard math Standard regularity conditions in Theorem 1: bounded logits, Lipschitz losses, finite first moments, and β∈[0,1].
    Listed before Theorem 1 to obtain the target-risk decomposition; these are conventional for generalization bounds.
  • standard math READOUT is permutation-invariant.
    Used in Eq. (4), following Xu et al. 2018a; standard for graph-level pooling.

pith-pipeline@v1.3.0-daily-deepseek · 14689 in / 11814 out tokens · 130494 ms · 2026-08-03T08:24:55.549868+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Cross-Resolution Semantic Learning for Graph Domain Adaptation." pith.science (2026). https://pith.science/paper/VZCVS6YU

@misc{pith2026260729365,
  author       = {Pith},
  title        = {Pith review of: Cross-Resolution Semantic Learning for Graph Domain Adaptation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VZCVS6YU}},
  note         = {Machine review of arXiv:2607.29365}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Graph Domain Adaptation (GDA) transfers predictive knowledge from labeled source graphs to unlabeled target graphs under distribution shift. Existing methods align representations or regularize graph structures, but do not explicitly model how class-discriminative knowledge learned at different source neighborhood ranges should be routed across target ranges. We call the neighborhood range encoded by a graph representation its propagation resolution and define semantic resolution shift as a cross-domain change in the propagation resolutions at which class-discriminative evidence is strongest. Such shifts can make fixed same-resolution pairing suboptimal and increase the risk of negative transfer. To address this issue, we propose Cross-Resolution Semantic Learning (CReSL), a GDA method that learns soft sourceto-target resolution correspondence from cross-domain class structure. First, CReSL constructs a multi-resolution representation bank using a shared Graph Neural Network and learnable resolution embeddings, with a resolution-indexed expert for each source resolution. Second, CReSL introduces Cross-Resolution Prototype Transport, which constructs class-resolution prototypes from source labels and soft target posteriors and converts cross-domain prototype discrepancies into expert-specific routing over target resolutions. Third, CReSL introduces Cross-Resolution Target Grafting, which constructs posterior-weighted target-to-source prototype displacements and enforces correspondence-weighted prediction consistency for instance-level adaptation under class uncertainty. Extensive experiments on graph benchmarks under diverse domain shifts show that CReSL outperforms strong representative baselines across most settings.

Figures

Figures reproduced from arXiv: 2607.29365 by Haoze Huang, Shangsong Liang, Yingxu Wang, Zhongkai Zheng.

Figure 1
Figure 1. Figure 1: Semantic resolution shift in GDA. (a) Informative [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Challenges in learning source-to-target resolution [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of CReSL. A shared GNN with learnable resolution embeddings constructs comparable resolution-specific [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: T-SNE visualizations of CReSL and the baseline [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Oracle compatibility and learned cross-resolution [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Visualization of cross-resolution target grafting on [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

126 extracted references · 16 linked inside Pith

  1. [1]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

    Dream: a dual variational framework for unsupervised graph domain adaptation , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

  2. [2]

    Deal: An unsupervised domain adaptive framework for graph-level classification , author=

  3. [3]

    2023 , organization=

    Coco: A coupled contrastive framework for unsupervised domain adaptive graph classification , author=. 2023 , organization=

    2023

  4. [4]

    Sa-gda: Spectral augmentation for graph domain adaptation , author=

  5. [5]

    arXiv preprint arXiv:2410.06883 , year=

    Degree-Conscious Spiking Graph for Cross-Domain Adaptation , author=. arXiv preprint arXiv:2410.06883 , year=

    arXiv

  6. [6]

    Advances in neural information processing systems , volume=

    Analysis of representations for domain adaptation , author=. Advances in neural information processing systems , volume=

  7. [7]

    arXiv preprint arXiv:2602.00656 , year=

    Riemannian Flow Matching for Disentangled Graph Domain Adaptation , author=. arXiv preprint arXiv:2602.00656 , year=

    Pith/arXiv arXiv

  8. [8]

    2010 , publisher=

    Hilbert space embeddings and metrics on probability measures , author=. 2010 , publisher=

    2010

  9. [9]

    Unsupervised domain adaptive graph convolutional networks , author=

  10. [10]

    Tfgda: Exploring topology and feature alignment in semi-supervised graph domain adaptation through robust clustering , author=

  11. [11]

    Towards multi-resolution spatiotemporal graph learning for medical time series classification , author=

  12. [12]

    IEEE Transactions on Geoscience and Remote Sensing , year=

    PixDiff: Multi-Resolution Diffusion Network with Pixelization for Hyperspectral Anomaly Detection , author=. IEEE Transactions on Geoscience and Remote Sensing , year=

  13. [13]

    Multi-scale graph learning for anti-sparse downscaling , author=

  14. [14]

    IEEE Transactions on Neural Networks and Learning Systems , volume=

    Cross-domain graph convolutions for adversarial unsupervised domain adaptation , author=. IEEE Transactions on Neural Networks and Learning Systems , volume=. 2021 , publisher=

    2021

  15. [15]

    Semi-supervised domain adaptation in graph transfer learning , author=

  16. [16]

    Machine learning , year=

    A theory of learning from different domains , author=. Machine learning , year=

  17. [17]

    IEEE Transactions on Information Theory , volume=

    Refinements of Pinsker's inequality , author=. IEEE Transactions on Information Theory , volume=. 2003 , publisher=

    2003

  18. [18]

    Foundations and Trends

    Computational optimal transport with applications to data sciences , author=. Foundations and Trends. 2019 , publisher=

    2019

  19. [19]

    2018 , publisher=

    Foundations of machine learning , author=. 2018 , publisher=

    2018

  20. [20]

    2023 , organization=

    Structural re-weighting improves graph domain adaptation , author=. 2023 , organization=

    2023

  21. [21]

    2024 , pages=

    Rethinking Propagation for Unsupervised Graph Domain Adaptation , author=. 2024 , pages=

    2024

  22. [22]

    Pairwise Alignment Improves Graph Domain Adaptation , author=

  23. [23]

    arXiv preprint arXiv:2502.06808 , year=

    On the benefits of attribute-driven graph domain adaptation , author=. arXiv preprint arXiv:2502.06808 , year=

    Pith/arXiv arXiv

  24. [24]

    Smoothness really matters: A simple yet effective approach for unsupervised graph domain adaptation , author=

  25. [25]

    arXiv preprint arXiv:2402.16374 , year=

    Graph learning under distribution shifts: A comprehensive survey on domain adaptation, out-of-distribution, and continual learning , author=. arXiv preprint arXiv:2402.16374 , year=

    Pith/arXiv arXiv

  26. [26]

    ACM Transactions on Knowledge Discovery from Data , volume=

    Graph domain adaptation: A generative view , author=. ACM Transactions on Knowledge Discovery from Data , volume=. 2024 , publisher=

    2024

  27. [27]

    Graph adaptive knowledge transfer for unsupervised domain adaptation , author=

  28. [28]

    IEEE Transactions on Knowledge and Data Engineering , volume=

    Graph transfer learning via adversarial domain adaptation with graph convolution , author=. IEEE Transactions on Knowledge and Data Engineering , volume=. 2022 , publisher=

    2022

  29. [29]

    Gcan: Graph convolutional adversarial network for unsupervised domain adaptation , author=

  30. [30]

    SPA: A graph spectral alignment perspective for domain adaptation , author=

  31. [31]

    Disentangled graph spectral domain adaptation , author=

  32. [32]

    IEEE Transactions on Artificial Intelligence , volume=

    Graph learning: A survey , author=. IEEE Transactions on Artificial Intelligence , volume=. 2021 , publisher=

    2021

  33. [33]

    ACM Transactions on Intelligent Systems and Technology , volume=

    A survey on graph representation learning methods , author=. ACM Transactions on Intelligent Systems and Technology , volume=. 2024 , publisher=

    2024

  34. [34]

    IEEE Transactions on Knowledge and Data Engineering , volume=

    Deep learning on graphs: A survey , author=. IEEE Transactions on Knowledge and Data Engineering , volume=. 2020 , publisher=

    2020

  35. [35]

    IEEE transactions on knowledge and data engineering , volume=

    Graph self-supervised learning: A survey , author=. IEEE transactions on knowledge and data engineering , volume=. 2022 , publisher=

    2022

  36. [36]

    IEEE transactions on neural networks and learning systems , volume=

    A comprehensive survey on graph neural networks , author=. IEEE transactions on neural networks and learning systems , volume=. 2020 , publisher=

    2020

  37. [37]

    Graph domain adaptation via theory-grounded spectral regularization , author=

  38. [38]

    Improving graph domain adaptation with network hierarchy , author=

  39. [39]

    arXiv preprint arXiv:2508.05182 , year=

    SPA++: Generalized Graph Spectral Alignment for Versatile Domain Adaptation , author=. arXiv preprint arXiv:2508.05182 , year=

    Pith/arXiv arXiv

  40. [40]

    Universal domain adaptation , author=

  41. [41]

    Neurocomputing , volume=

    Deep visual domain adaptation: A survey , author=. Neurocomputing , volume=. 2018 , publisher=

    2018

  42. [42]

    2015 , organization=

    Unsupervised domain adaptation by backpropagation , author=. 2015 , organization=

    2015

  43. [43]

    arXiv preprint arXiv:1710.06924 , year=

    Visda: The visual domain adaptation challenge , author=. arXiv preprint arXiv:1710.06924 , year=

    Pith/arXiv arXiv

  44. [44]

    ACM Transactions on Intelligent Systems and Technology (TIST) , volume=

    A survey of unsupervised deep domain adaptation , author=. ACM Transactions on Intelligent Systems and Technology (TIST) , volume=. 2020 , publisher=

    2020

  45. [45]

    2019 , organization=

    Bridging theory and algorithm for domain adaptation , author=. 2019 , organization=

    2019

  46. [46]

    Can modifying data address graph domain adaptation? , author=

  47. [47]

    Revisiting, benchmarking and understanding unsupervised graph domain adaptation , author=

  48. [48]

    IEEE transactions on neural networks and learning systems , volume=

    On representation knowledge distillation for graph neural networks , author=. IEEE transactions on neural networks and learning systems , volume=. 2022 , publisher=

    2022

  49. [49]

    Mugsi: Distilling gnns with multi-granularity structural information for graph classification , author=

  50. [50]

    IEEE Transactions on Knowledge and Data Engineering , volume=

    A teacher-free graph knowledge distillation framework with dual self-distillation , author=. IEEE Transactions on Knowledge and Data Engineering , volume=. 2024 , publisher=

    2024

  51. [51]

    Label attentive distillation for GNN-based graph classification , author=

  52. [52]

    Adagmlp: Adaboosting gnn-to-mlp knowledge distillation , author=

  53. [53]

    Simple yet effective graph distillation via clustering , author=

  54. [54]

    ACM Computing Surveys , volume=

    Knowledge distillation on graphs: A survey , author=. ACM Computing Surveys , volume=. 2025 , publisher=

    2025

  55. [55]

    IEEE Transactions on Knowledge and Data Engineering , volume=

    Graph condensation: A survey , author=. IEEE Transactions on Knowledge and Data Engineering , volume=. 2025 , publisher=

    2025

  56. [56]

    Self-supervised learning for graph dataset condensation , author=

  57. [57]

    Condensing graphs via one-step gradient matching , author=

  58. [58]

    Hogda: Boosting semi-supervised graph domain adaptation via high-order structure-guided adaptive feature alignment , author=

  59. [59]

    IEEE Transactions on Image Processing , volume=

    Attention regularized Laplace graph for domain adaptation , author=. IEEE Transactions on Image Processing , volume=. 2022 , publisher=

    2022

  60. [60]

    Pattern Recognition , volume=

    Graph domain adaptation with localized graph signal representations , author=. Pattern Recognition , volume=. 2024 , publisher=

    2024

  61. [61]

    Structure-Attribute Transformations with Markov Chain Boost Graph Domain Adaptation , author=

  62. [62]

    Coupling category alignment for graph domain adaptation , author=

  63. [63]

    arXiv preprint arXiv:2602.10489 , year=

    Learning Adaptive Distribution Alignment with Neural Characteristic Function for Graph Domain Adaptation , author=. arXiv preprint arXiv:2602.10489 , year=

    arXiv

  64. [64]

    arXiv preprint arXiv:2602.10506 , year=

    Learning Structure-Semantic Evolution Trajectories for Graph Domain Adaptation , author=. arXiv preprint arXiv:2602.10506 , year=

    arXiv

  65. [65]

    T2-gnn: Graph neural networks for graphs with incomplete features and structure via teacher-student distillation , author=

  66. [66]

    arXiv preprint arXiv:2310.09202 , year=

    Graph distillation with eigenbasis matching , author=. arXiv preprint arXiv:2310.09202 , year=

    Pith/arXiv arXiv

  67. [67]

    arXiv preprint arXiv:2110.08727 , year=

    Graph-less neural networks: Teaching old mlps new tricks via distillation , author=. arXiv preprint arXiv:2110.08727 , year=

    Pith/arXiv arXiv

  68. [68]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=

    A comprehensive survey of dataset distillation , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=. 2023 , publisher=

    2023

  69. [69]

    Stability of graph scattering transforms , author=

  70. [70]

    Journal of Machine Learning Research , volume=

    Transferability of spectral graph convolutional neural networks , author=. Journal of Machine Learning Research , volume=

  71. [71]

    2021 , organization=

    Interpretable stability bounds for spectral graph filters , author=. 2021 , organization=

    2021

  72. [72]

    2015 ieee information theory workshop (itw) , pages=

    Deep learning and the information bottleneck principle , author=. 2015 ieee information theory workshop (itw) , pages=. 2015 , organization=

    2015

  73. [73]

    Advances in applied probability , volume=

    Integral probability metrics and their generating classes of functions , author=. Advances in applied probability , volume=. 1997 , publisher=

    1997

  74. [74]

    Information-theoretic analysis of generalization capability of learning algorithms , author=

  75. [75]

    2020 , organization=

    Stability of graph neural networks to relative perturbations , author=. 2020 , organization=

    2020

  76. [76]

    Graph domain adaptation with dual-branch encoder and two-level alignment for whole slide image-based survival prediction , author=

  77. [77]

    Higda: Hierarchical graph of nodes to learn local-to-global topology for semi-supervised domain adaptation , author=

  78. [78]

    Joint European Conference on Machine Learning and Knowledge Discovery in Databases , pages=

    Theoretical analysis of domain adaptation with optimal transport , author=. Joint European Conference on Machine Learning and Knowledge Discovery in Databases , pages=. 2017 , organization=

    2017

  79. [79]

    Communications of the ACM , volume=

    A theory of the learnable , author=. Communications of the ACM , volume=. 1984 , publisher=

    1984

  80. [80]

    Journal of machine learning research , volume=

    Rademacher and gaussian complexities: Risk bounds and structural results , author=. Journal of machine learning research , volume=

Showing first 80 references.